Tower drum form monitoring method based on remote sensing image feature analysis

Abstract

The invention provides a tower shape monitoring method based on remote sensing image feature analysis, and relates to the field of image recognition, dynamic frequency separation is carried out through interference and phase superposition of different frequency components, and separation of high-frequency and low-frequency components can be dynamically adjusted according to different contents of an image; a multi-dimensional interaction enhancement module is designed by combining the principles of a lattice structure model and a turbulence model, and high-frequency and low-frequency information is enhanced through dynamic interaction and fused with multi-scale features; according to the method, a self-adaptive segmentation framework is constructed by combining local adaptability enhancement and a self-organizing criticality theory in a biological vision mechanism, so that a segmentation result is finer and adapts to different tower drum remote sensing image contents, a tower drum region can be extracted more accurately, and a better segmentation effect is shown in a complex environment.

Description

Technical Field

[0001] The present invention belongs to the field of image recognition, and in particular relates to a tower morphology monitoring method based on remote sensing image feature analysis. Background Art

[0002] Tower morphology monitoring plays an important role in modern engineering structure monitoring, especially in the safety assessment of tower structures such as communication towers, wind turbine towers, and transmission towers. As the tower structure ages and is affected by external environmental factors (such as wind, temperature differences, humidity, and earthquakes), the tower may undergo morphological changes such as tilt, cracks, corrosion, and local deformation. These changes may affect the stability and safety of the tower. Therefore, regular and accurate tower morphology monitoring is key to ensuring its structural health and safety. Traditional tower morphology monitoring methods rely on manual inspections or periodic physical testing, which often have certain limitations. Manual inspections are not only time-consuming and limited by the capabilities of the inspectors and external environmental conditions, but also fail to monitor the tower status in real time and comprehensively, which can easily lead to the omission of certain potential problems. Monitoring through physical sensors, while able to provide a certain degree of real-time data, has high deployment and maintenance costs and cannot provide global and fine-grained status information of the tower surface.

[0003] In recent years, the development of remote sensing technology and computer vision technology has made tower morphology monitoring methods based on remote sensing images an emerging solution. Remote sensing images can be taken by satellite or drone to achieve all-round, high-resolution image acquisition of tower structures, and can cover a wide area, reducing the manual intervention and detection costs in traditional methods. Remote sensing image feature analysis technology, especially image processing and analysis methods based on computer vision, provides new ideas for tower morphology monitoring. Through remote sensing images, various features of the tower surface and morphology (such as texture, cracks, rust, etc.) can be extracted, and these features can be automatically identified and classified through deep learning models, thereby achieving efficient monitoring of tower morphology changes. Furthermore, the use of image enhancement, multi-scale feature extraction, frequency domain analysis and other technologies can more accurately detect minor deformations, cracks and other problems, thereby improving the accuracy and reliability of monitoring. Summary of the Invention

[0004] The present invention provides a tower morphology monitoring method based on remote sensing image feature analysis. Dynamic frequency separation is performed through interference and phase superposition of different frequency components. The separation of high-frequency and low-frequency components can be dynamically adjusted according to the different contents of the image, thereby more accurately capturing the details and global semantic information of the image. By combining the principles of lattice structure model and turbulence model, a multi-dimensional interactive enhancement module is designed. High-frequency and low-frequency information are enhanced through dynamic interaction and fused with multi-scale features. The detailed information of the tower morphology in the remote sensing image can be more accurately enhanced while maintaining the stability and accuracy of the global semantics. An adaptive segmentation framework is constructed by combining local adaptive enhancement and self-organized criticality theory in biological vision mechanism. The segmentation result is more refined and adaptable to the contents of different tower remote sensing images, which can more accurately extract the tower area and show good segmentation effect in complex environments.

[0005] In order to achieve the above-mentioned purpose, the present invention provides the following technical solution: a tower morphology monitoring method based on remote sensing image feature analysis, comprising the following steps.

[0006] S1. Traditional convolution operations cannot adaptively adjust the feature contribution of each channel according to the image content. Channel adaptive feature mapping is designed to improve the initial feature extraction of tower remote sensing images by adaptively learning the importance of each channel.

[0007] S2. Introduce a multi-frequency interference mechanism to perform dynamic frequency separation through interference and phase superposition of different frequency components, improving the traditional static separation of high and low frequencies through simple fast Fourier transform.

[0008] S3. Combining the lattice structure model and turbulence model, a multi-dimensional interaction enhancement module is designed to enhance high-frequency and low-frequency information through dynamic interaction, and to enhance the multi-level details of the tower remote sensing image by simulating the vortex effect through multi-scale feature exchange.

[0009] S4. We introduce the visual processing of the cat's eye system and the theory of self-organized criticality in complex systems, and construct an adaptive segmentation framework through dynamic adjustment and adaptive segmentation of image content.

[0010] S5. Construct a tower morphology monitoring model, enhance the features of tower remote sensing images through dynamic frequency separation and multi-dimensional interactive enhancement modules, select the region of interest in the tower remote sensing image through image segmentation and region extraction, and input the segmented region into the classifier to classify the tower morphology.

[0011] Preferably, in step S1, the tower remote sensing image is first preprocessed, and then the initial features of the tower remote sensing image are extracted. The specific steps are:

[0012] S11. Preprocess the tower remote sensing images. First, use adaptive noise suppression technology, combined with Gaussian filtering and change detection of local image features, to remove environmental interference and sensor noise. Then, use multispectral fusion to perform a weighted combination of visible and near-infrared remote sensing images to enhance the tower's structural features and texture details. Adaptive light balancing is used to address uneven brightness caused by varying lighting conditions.

[0013] S12, input the pre-processed tower remote sensing image F in ∈R H×W×C In the initial feature extraction block, R is a real number domain, H, W, and C are height, width, and total number of channels respectively. Initial feature extraction is performed by adaptively weighting the tower remote sensing image features to highlight the tower features. The initial feature extraction process of the tower remote sensing image is as follows:

[0014]

[0015] Where, I in ∈R H×W×C is the initial feature of the tower remote sensing image, W αi is the dynamically generated weighting coefficient, is the feature of the i-th channel at (p, q).

[0016] Preferably, in step S1, the input tower remote sensing image is preprocessed and initial feature extraction is performed. The tower remote sensing image is preprocessed by adaptive noise suppression technology, multispectral fusion and adaptive illumination equalization to make the tower morphological features clearer and more detailed, providing high-quality input images for subsequent feature extraction and morphological monitoring. The tower morphology in the remote sensing image may have different features due to changes in terrain, weather, etc. Therefore, it is necessary to adjust the contribution of each image channel according to the local features of the tower remote sensing image to improve the accuracy of tower morphological extraction; initial feature extraction is performed by adaptive weighting, and the features of each channel are adaptively weighted to ensure that the tower morphological information is highlighted and the background interference information is suppressed, thereby improving the tower morphological monitoring accuracy.

[0017] Preferably, in step S2, the multi-frequency interference mechanism comprises the following steps:

[0018] S21. Input the initial features of the tower remote sensing image I in ∈R H×W×c Each pixel of the tower remote sensing image contains the interference of multiple frequency components. Each frequency component has different phase information. The complex feature map in the frequency domain is calculated by frequency components and phase information. The interference expression model of the frequency components of the tower remote sensing image is as follows:

[0019]

[0020] Where, F interfere (u, v) is the complex feature map in the frequency domain, (u, v) is the coordinate in the frequency domain, K is the number of frequency components, A k is the amplitude of the kth frequency component, f (k，x) and f (k，y) are the frequency components of the kth frequency component along the x and y directions in the frequency domain, θ k is the phase of the kth frequency component;

[0021] S22, design weight factor ω k The high-frequency and low-frequency components are calculated by the dynamic frequency separation mechanism, and the weight factor ω k The specific calculation formula is:

[0022]

[0023] Where σ is the sigmoid function, For I in Local texture features, LW is I in The global semantic features of k The frequency domain complex feature map F interfere (u, v) is divided into high-frequency information and low-frequency information

[0024]

[0025] Preferably, in step S2, traditional frequency separation usually performs static separation of high-frequency information and low-frequency information through a simple fast Fourier transform. However, this method ignores the interaction effect between different frequency components; dynamic frequency separation is performed through interference and phase superposition of different frequency components, which can dynamically adjust the separation of high-frequency and low-frequency components according to the different contents of the image, thereby more accurately capturing the details and global semantic information of the image, and can enhance high-frequency components in areas with rich details and highlight low-frequency components in areas with simpler structures.

[0026] Preferably, in step S3, the multi-dimensional interaction enhancement module specifically comprises the following steps:

[0027] S31. Input high-frequency information and low-frequency information In the multi-dimensional interaction enhancement module, we first design an adaptive lattice interaction mechanism by referring to the lattice structure theory. The adaptive interaction coefficient is used to adjust the interaction between high-frequency and low-frequency features. The calculation steps of the adaptive lattice interaction mechanism are as follows:

[0028]

[0029] Where, and are the high-frequency information and low-frequency information after interaction, α inter is the adaptive interaction coefficient, which is calculated by the strength of the local gradient inter , controls the interaction strength between high-frequency and low-frequency components, and the adaptive interaction coefficient α inter The specific calculation formula is:

[0030]

[0031] Where, β inter In order to control the sensitivity of the weighting coefficient to the local gradient, G local is the local gradient value, γ inter is the bias term, which controls the threshold of the gradient value;

[0032] S32. Introducing a turbulence model into the multi-dimensional interaction enhancement module to simulate the interaction of features at different scales of the image, wherein the turbulence model is:

[0033]

[0034] Where, I mul is the multi-scale interaction feature, L is the number of scale layers, α l is the weighting coefficient for each scale, S l is the characteristic intensity related to the scale of the first layer, I l is the feature at the lth level, β l is the parameter that controls the intensity of the turbulence model, is the Laplace operator;

[0035] S33. After completing the interaction enhancement of high frequency and low frequency and the interaction of different scale features, the different frequency and multi-scale interaction features are fused to obtain the multi-dimensional interaction feature I inter-mul .

[0036] Preferably, in step S3, a multi-dimensional interactive enhancement module is designed by combining the principles of lattice structure model and turbulence model. This module not only enhances high-frequency and low-frequency information through dynamic interaction, but also enhances the multi-level details of the image through multi-scale feature exchange and imitating the vortex effect. Compared with traditional methods, this design can more accurately enhance the detailed information of the tower shape in remote sensing images while maintaining the stability and accuracy of global semantics. It is particularly suitable for application in complex remote sensing image feature analysis.

[0037] Preferably, in step S4, the specific steps of the adaptive segmentation framework are:

[0038] S41. By introducing the local contrast enhancement mechanism in the cat's eye system, the important areas in the image are dynamically enhanced. The local areas in the tower remote sensing image are highlighted by an enhancement model based on local contrast. The contrast enhancement formula is:

[0039]

[0040] Where, I enha (x, y) is the pixel value after enhancement at position (x, y), I inter-mul (x, y) is the multidimensional interaction feature I inter-mul The pixel value at (x, y), I lmean (x, y) is the mean of the local area, I lstd (x, y) is the standard deviation of the local area;

[0041] S42. Introduce the self-organized criticality theory to dynamically adjust the segmentation threshold. Automatically adjust the segmentation threshold according to the gradient change of the image. The segmentation threshold T(ti) is calculated as follows:

[0042]

[0043] Where, T init is the initial segmentation threshold, α T To control the rate of dynamic adjustment of the threshold, I lstd (x, y, ti) is the standard deviation of the local area at time ti, indicating the local variation of the image;

[0044] S43, extracting the region of interest in the tower remote sensing image by calculating the relationship between the difference between the enhanced pixel value and the segmentation domain pixel value and the segmentation threshold, wherein the region boundary detection is:

[0045]

[0046] Where, if Region is 1, the pixel is the target area; if Region is 0, the pixel is the non-target area. seed The starting value of the segmentation domain.

[0047] Preferably, in step S4, the image segmentation and region extraction method proposed in the present invention combines the local adaptive enhancement in the biological vision mechanism and the dynamic segmentation threshold adjustment of the self-organizing criticality theory. By imitating the cat's eye vision mechanism and adopting adaptive contrast enhancement, the detail performance of the local area of the image is improved; by introducing the self-organizing criticality model, the segmentation threshold of the image is dynamically adjusted, so that the segmentation result is more refined and adapted to the content of different tower remote sensing images. This method can extract the tower area more accurately and show better segmentation effect in complex environments.

[0048] Preferably, in step S5, the tower shape monitoring model comprises the following specific steps:

[0049] S51. First, the tower remote sensing image is input into the tower morphology monitoring model. The tower remote sensing image is preprocessed by removing environmental interference and sensor noise through adaptive noise suppression technology and multispectral fusion processing. Next, the adaptive weighted feature extraction method is used to dynamically weight each channel of the tower remote sensing image to extract the initial features of the tower remote sensing image, providing high-quality input features for subsequent tower remote sensing image feature extraction.

[0050] S52. Inputting the initial features of the tower remote sensing image into M information interaction modules composed of a multi-frequency interference mechanism and a multi-dimensional interaction enhancement module, wherein the multi-frequency interference mechanism is used to perform dynamic frequency separation, and the separation of high-frequency information and low-frequency information is dynamically adjusted through phase superposition and interference of frequency components to obtain high-frequency information and low-frequency information of the tower remote sensing image. Subsequently, the multi-dimensional interaction enhancement module enhances the interaction between high-frequency information and low-frequency information and the fusion of multi-scale features by combining a lattice structure model and a turbulence model. Finally, the M interaction features are fused to obtain enhanced tower remote sensing image features.

[0051] S53. After completing feature enhancement, the model uses image segmentation and region extraction methods, combined with the local contrast enhancement mechanism and self-organized criticality theory in the cat's eye system, to dynamically adjust the segmentation threshold and accurately extract the region of interest of the tower. Finally, the extracted tower remote sensing area is input into the classifier for tower morphology classification. The tower status is classified according to different features in the area, including surface cracks, tilt and deformation, material aging, local deformation, and normal monitoring.

[0052] Preferably, in step S5, the tower morphology monitoring model can accurately and automatically identify tower morphology changes, such as cracks, tilt, and material aging, through innovative tower remote sensing image processing technology, significantly improving the tower image quality and detail performance. The model combines multi-frequency interference mechanism and adaptive frequency separation to enhance the interaction of high-frequency and low-frequency information. At the same time, through adaptive segmentation and region extraction technology, it ensures that the tower's area of interest is accurately extracted. The model also has long-term monitoring and trend prediction functions, which helps to identify potential risks in advance, reduce manual intervention, and improve the safety, reliability and maintenance efficiency of the tower.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] The present invention provides a tower morphology monitoring method based on remote sensing image feature analysis, which improves the traditional static separation of high-frequency information and low-frequency information after simple fast Fourier transform, and performs dynamic frequency separation through interference and phase superposition of different frequency components. The method can dynamically adjust the separation of high-frequency and low-frequency components according to the different contents of the image, thereby more accurately capturing the details and global semantic information of the image; by combining the principles of lattice structure model and turbulence model, a multi-dimensional interactive enhancement module is designed, and high-frequency and low-frequency information are enhanced through dynamic interaction and integrated with multi-scale features. Compared with traditional methods, this design can more accurately enhance the detailed information of the tower morphology in the remote sensing image while maintaining the stability and accuracy of the global semantics, and is particularly suitable for application in complex remote sensing image feature analysis; combining local adaptability enhancement and self-organized criticality theory in biological vision mechanism to construct an adaptive segmentation framework, so that the segmentation result is more refined and adapts to the contents of different tower remote sensing images, can more accurately extract the tower area, and show good segmentation effect in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a flow chart of a tower morphology monitoring method based on remote sensing image feature analysis provided by the present invention.

[0056] Figure 2 This is a structural diagram of the multi-frequency interference mechanism provided by the present invention.

[0057] Figure 3 This is a structural diagram of the multi-dimensional interaction enhancement module provided by the present invention.

[0058] Figure 4 This is a structural diagram of the tower morphology monitoring model provided by the present invention.

[0059] Figure 5 This is a tower remote sensing image detection effect diagram provided by the present invention. DETAILED DESCRIPTION

[0060] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0061] See also Figures 1 to 5The present invention provides a tower morphology monitoring method based on remote sensing image feature analysis, which improves the traditional static separation of high-frequency information and low-frequency information after simple fast Fourier transform, and performs dynamic frequency separation through interference and phase superposition of different frequency components. It can dynamically adjust the separation of high-frequency and low-frequency components according to the different contents of the tower remote sensing image, so as to more accurately capture the details of the image and global semantic information; by combining the principles of lattice structure model and turbulence model, a multi-dimensional interactive enhancement module is designed, which enhances high-frequency and low-frequency information through dynamic interaction and fuses them with multi-scale features. Compared with traditional methods, this design can more accurately enhance the detailed information of the tower morphology in the remote sensing image while maintaining the stability and accuracy of the global semantics, and is particularly suitable for application in complex remote sensing image feature analysis; combining local adaptability enhancement and self-organized criticality theory in biological vision mechanism to construct an adaptive segmentation framework, so that the segmentation result is more refined and adapts to the contents of different tower remote sensing images, can more accurately extract the tower area, and show better segmentation effect in complex environments.

[0062] See Figure 1 As shown, a tower morphology monitoring method based on remote sensing image feature analysis in an embodiment of the present application.

[0063] S1. Traditional convolution operations cannot adaptively adjust the feature contribution of each channel according to the image content. Channel adaptive feature mapping is designed to improve the initial feature extraction of tower remote sensing images by adaptively learning the importance of each channel.

[0064] Furthermore, the tower remote sensing image is first preprocessed, and then the initial features of the tower remote sensing image are extracted. The specific steps are as follows.

[0065] S11. Preprocess the tower remote sensing images. First, use adaptive noise suppression technology combined with Gaussian filtering and change detection of local image features to remove environmental interference and sensor noise. Then, use multispectral fusion methods to perform a weighted combination of remote sensing images in the visible light and near-infrared bands to enhance the structural features and texture details of the tower. At the same time, adaptive light balancing is used to solve the problem of uneven brightness caused by different lighting conditions.

[0066] S12, input the pre-processed tower remote sensing image F in ∈R H×W×C Go to the initial feature extraction block, R is the real number domain, H, W, and C are the height, width, and total number of channels respectively, which are set to 640, 640, and 3. Initial feature extraction is performed through adaptive weighting of tower remote sensing image features to highlight tower features. The initial feature extraction process of tower remote sensing image is as follows:

[0067]

[0068] Where, I in ∈R H×W×C is the initial feature of the tower remote sensing image, W αi is a dynamically generated weighting coefficient, with a value range of W αi ∈[0, 1], the initial value is 0.3, and it is dynamically adjusted according to the local features of the image. is the feature of the i-th channel at (p, q), is the spatial importance map, which indicates the importance of the channel at each position.

[0069] S2. Introduce a multi-frequency interference mechanism to perform dynamic frequency separation through interference and phase superposition of different frequency components, improving the traditional static separation of high and low frequencies through simple fast Fourier transform.

[0070] Further, if Figure 2 As shown, the specific steps of the multi-frequency interference mechanism are as follows.

[0071] S21. Input the initial features of the tower remote sensing image I in ∈R H×W×c Each pixel of the tower remote sensing image contains the interference of multiple frequency components. Each frequency component has different phase information. The complex feature map in the frequency domain is calculated by frequency components and phase information. The interference expression model of the frequency components of the tower remote sensing image is as follows:

[0072]

[0073] Where, F interfere (u, v) is the complex feature map in the frequency domain, (u, v) is the coordinate in the frequency domain, K is the number of frequency components, K = 204800, A k is the amplitude of the kth frequency component, indicating the intensity of the frequency component, and its value range is A k ∈[0, 1], with an initial value of 0.3, represents the frequency component of the image, which is obtained by superimposing multiple frequency components and reflects the interference and superposition of different frequency components in the image. (k，x) and f (k，y) are the frequency components of the kth frequency component along the x and y directions in the frequency domain, which determine the position of the frequency component in the frequency domain, θ k is the phase of the kth frequency component, which controls the relative timing of the frequency components and has a value range of θ k ∈[0, 2π], the initial value is π.

[0074] S22. In order to extract high-frequency and low-frequency information, design the weight factor ω k The high-frequency and low-frequency components are calculated by the dynamic frequency separation mechanism, and the weight factor ω k The specific calculation formula is:

[0075]

[0076] Where σ is the sigmoid function, For I in The local texture features are extracted using local gradients, and LW is I in The global semantic features of image I in Perform low-pass filtering to extract global structural information, and use the weight factor ω k The frequency domain complex feature map F interfere (u,v) is divided into high-frequency information and low-frequency information,

[0077]

[0078] S3. Combining the lattice structure model and turbulence model, a multi-dimensional interaction enhancement module is designed to enhance high-frequency and low-frequency information through dynamic interaction, and to enhance the multi-level details of the tower remote sensing image by simulating the vortex effect through multi-scale feature exchange.

[0079] Further, if Figure 3 As shown, the specific steps of the multi-dimensional interaction enhancement module are as follows.

[0080] S31. Input high-frequency information and low-frequency information In the multi-dimensional interaction enhancement module, we first design an adaptive lattice interaction mechanism by referring to the lattice structure theory. The adaptive interaction coefficient is used to adjust the interaction between high-frequency and low-frequency features. The calculation steps of the adaptive lattice interaction mechanism are as follows:

[0081]

[0082] Where, and are the high-frequency information and low-frequency information after interaction, α inter is the adaptive interaction coefficient, which is calculated by the strength of the local gradient inter , controls the interaction strength between high-frequency and low-frequency components, and its value range is α inter ∈[0, 1], the initial value is 0.5, the adaptive interaction coefficient α inter The specific calculation formula is:

[0083]

[0084] Where, β inter In order to control the sensitivity of the weighting coefficient to the local gradient, the value range is β inter ∈[1, 5], the initial value is 2, Glocal is the local gradient value, γ inter is the bias term, which controls the threshold of the gradient value and has a value range of γ inter ∈[0, 1], the initial value is 0.3, when the details of the tower remote sensing image area are richer (the local gradient is larger), the adaptive interaction coefficient α inter The larger the value, the higher the high-frequency components will be. On the contrary, when the details of the tower remote sensing image area are less, the weight of the low-frequency components will be greater.

[0085] S32. Introducing a turbulence model into the multi-dimensional interaction enhancement module to simulate the interaction of features at different scales in the image, further improving the representation of image details and semantics, the turbulence model is:

[0086]

[0087] Where, I mul is a multi-scale interactive feature, L is the number of scale layers, the value range is L∈[5,7], the initial value is 5, α l The weighting coefficient for each scale controls the contribution strength of each scale feature. S l is the characteristic strength related to the scale of the first layer, and its value range is α l ∈[0, 1], the initial value is 0.3, I l is the feature at the lth level, β l It is a parameter that controls the intensity of the turbulence model and affects the intensity of the characteristic exchange between scales. Its value range is β l ∈[0, 5], the initial value is 2, is the Laplace operator, which is used to represent the diffusion degree of the vortex structure in the image and can capture the interaction between different scales.

[0088] S33. After completing the interaction enhancement of high frequency and low frequency and the interaction of different scale features, convolution and feature fusion are used to fuse the different frequency and multi-scale interaction features to obtain the multi-dimensional interaction feature I inter-mul , the fusion process is:

[0089]

[0090] Where, Conv 1×1 It is a 1×1 convolution operation, and Concat is a fusion operation.

[0091] S4. We introduce the visual processing of the cat's eye system and the theory of self-organized criticality in complex systems, and construct an adaptive segmentation framework through dynamic adjustment and adaptive segmentation of image content.

[0092] Furthermore, the specific steps of the adaptive segmentation framework are as follows.

[0093] S41. By introducing the local contrast enhancement mechanism in the cat's eye system, the important areas in the image are dynamically enhanced. The local areas in the tower remote sensing image are highlighted by an enhancement model based on local contrast. The contrast enhancement formula is:

[0094]

[0095] Where, I enha (x, y) is the pixel value after enhancement at position (x, y), I inter-mul (x, y) is the multidimensional interaction feature I inter-mul The pixel value at (x, y), I lmean (x, y) is the mean of the local area, N is the number of pixels in the local window, N = (2t + 1) 2 , t is the window radius, the range is t∈[1,3], the initial value is 1, I lstd (x, y) is the standard deviation of the local area,

[0096]

[0097] S42. The self-organized criticality theory is introduced to dynamically adjust the segmentation threshold, making the segmentation in different areas more refined. The threshold can be adaptively adjusted in important feature areas in the image (such as the edge of the tower), and the segmentation threshold is automatically adjusted according to the gradient change of the image. The segmentation threshold T(ti) is calculated as follows:

[0098]

[0099] Where, T init is the initial segmentation threshold, the initial value is set to 60, α T To control the rate of dynamic adjustment of the threshold, the value range is α T ∈[0, 1], the initial value is 0.5, I lstd (x, y, t i ) is the standard deviation of the local area at time ti, which represents the local change amplitude of the image.

[0100] S43, extracting the region of interest in the tower remote sensing image by calculating the relationship between the difference between the enhanced pixel value and the segmentation domain pixel value and the segmentation threshold, wherein the region boundary detection is:

[0101]

[0102] Where, if Region is 1, the pixel is the target area; if Region is 0, the pixel is the non-target area. seedis the starting value of the segmentation domain, the value range is t∈[100, 130], and the initial value is 110.

[0103] S5. Construct a tower morphology monitoring model, enhance the features of tower remote sensing images through dynamic frequency separation and multi-dimensional interactive enhancement modules, select the region of interest in the tower remote sensing image through image segmentation and region extraction, and input the segmented region into the classifier to classify the tower morphology.

[0104] Further, if Figure 4 As shown in the figure, the specific steps of the tower shape monitoring model are as follows.

[0105] S51. First, input the tower remote sensing image into the tower morphology monitoring model, pre-process the tower remote sensing image, and remove environmental interference and sensor noise through adaptive noise suppression technology and multi-spectral fusion processing to ensure the image quality. Next, use the adaptive weighted feature extraction method to dynamically weight each channel of the tower remote sensing image to extract the initial features of the tower remote sensing image, highlight the morphological features of the tower, suppress background interference, and provide high-quality input features for subsequent tower remote sensing image feature extraction.

[0106] S52. Input the initial features of the tower remote sensing image into M information interaction modules composed of a multi-frequency interference mechanism and a multi-dimensional interaction enhancement module, wherein the multi-frequency interference mechanism is used for dynamic frequency separation, and the separation of high-frequency information and low-frequency information is dynamically adjusted by phase superposition and interference of frequency components to obtain the high-frequency information and low-frequency information of the tower remote sensing image, so that the detail information and global semantics of the image can be more accurately reflected. Subsequently, the multi-dimensional interaction enhancement module enhances the interaction between high-frequency information and low-frequency information and the fusion of multi-scale features by combining the lattice structure model and the turbulence model, further improving the image detail expression and semantic stability. Finally, the M interaction features are fused to obtain the enhanced tower remote sensing image features.

[0107] S53. After completing feature enhancement, the model uses image segmentation and region extraction methods, combined with the local contrast enhancement mechanism and self-organized criticality theory in the cat's eye system, to dynamically adjust the segmentation threshold, accurately extract the area of interest of the tower, and highlight key details such as deformation and cracks of the tower. Finally, the extracted tower remote sensing area is input into the classifier for tower morphology classification, and the tower status is classified according to different features in the area, including surface cracks, tilt and deformation, material aging, local deformation, and normal monitoring.

[0108] Furthermore, in step S5, the tower morphology monitoring model is implemented based on the Pytorch framework through the Pycharm application, and the tower remote sensing image dataset is used to train the model and the test set is used to test it.

[0109] Further, if Figure 5 As shown in the figure, the remote sensing image of the tower shows normal tower shape monitoring results.

[0110] The above are only preferred embodiments of the present invention. It should be pointed out that those skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A tower morphology monitoring method based on remote sensing image feature analysis, characterized in that: The following steps are involved: S1. Traditional convolution operations cannot adaptively adjust the feature contribution of each channel according to the image content. Channel adaptive feature mapping is designed to improve the initial feature extraction of tower remote sensing images by adaptively learning the importance of each channel. S2. Introducing a multi-frequency interference mechanism to perform dynamic frequency separation through the interference and phase superposition of different frequency components, improving the traditional static separation of high and low frequencies through simple fast Fourier transform; S3. Design a multi-dimensional interactive enhancement module by combining the lattice structure model and the turbulence model. This module enhances high-frequency and low-frequency information through dynamic interaction and simulates the vortex effect through multi-scale feature exchange to enhance the multi-level details of the tower remote sensing image. S4. Introducing the visual processing of the cat's eye system and the theory of self-organized criticality in complex systems, we construct an adaptive segmentation framework through dynamic adjustment and adaptive segmentation of image content. S5. Build a tower morphology monitoring model, enhance the features of tower remote sensing images through dynamic frequency separation and multidimensional interactive enhancement modules, select the region of interest in the tower remote sensing image through an adaptive segmentation framework, and input the segmented region into the classifier to classify the tower morphology.

2. The tower morphology monitoring method based on remote sensing image feature analysis according to claim 1 is characterized in that: In step S1, the tower remote sensing image is first preprocessed, and then the initial features of the tower remote sensing image are extracted. The specific steps are as follows: S11. Preprocess the tower remote sensing images. First, use adaptive noise suppression technology, combined with Gaussian filtering and change detection of local image features, to remove environmental interference and sensor noise. Then, use multispectral fusion to perform a weighted combination of visible and near-infrared remote sensing images to enhance the tower's structural features and texture details. Adaptive light balancing is used to address uneven brightness caused by varying lighting conditions. S12, input the pre-processed tower remote sensing image F in ∈R H×W×C To the initial feature extraction block, R is the real number domain, H, W, and C are the height, width, and total number of channels respectively, and the initial feature extraction is performed through the adaptive weighting of the tower remote sensing image features. The initial feature extraction process of the tower remote sensing image is: Where, I in ∈R H×W×C is the initial feature of the tower remote sensing image, W αi is the dynamically generated weighting coefficient, is the feature of the i-th channel at (p, q).

3. The tower morphology monitoring method based on remote sensing image feature analysis according to claim 2 is characterized in that: In the step S2, the specific steps of the multi-frequency interference mechanism are as follows: S21. Input the initial features of the tower remote sensing image I in ∈R H×W×C Each pixel of the tower remote sensing image contains the interference of multiple frequency components. Each frequency component has different phase information. The complex feature map in the frequency domain is calculated by frequency components and phase information. The interference expression model of the frequency components of the tower remote sensing image is as follows: Where, F interfere (u, v) is the complex feature map in the frequency domain, (u, v) is the coordinate in the frequency domain, K is the number of frequency components, A k is the amplitude of the kth frequency component, f (k，x) and f (k，y) are the frequency components of the kth frequency component along the x and y directions in the frequency domain, θ k is the phase of the kth frequency component; S22, design weight factor ω k The high-frequency and low-frequency components are calculated by the dynamic frequency separation mechanism, and the weight factor ω k The specific calculation formula is: Where σ is the sigmoid function, For I in Local texture features, LW is I in The global semantic features of k The frequency domain complex feature map F interfere (u,v) is divided into high-frequency information and low-frequency information 4. The tower morphology monitoring method based on remote sensing image feature analysis according to claim 3 is characterized in that: In the S3 step, the multi-dimensional interaction enhancement module specifically comprises the following steps: S31. Input high-frequency information and low-frequency information In the multi-dimensional interaction enhancement module, we first design an adaptive lattice interaction mechanism by referring to the lattice structure theory. The adaptive interaction coefficient is used to adjust the interaction between high-frequency and low-frequency features. The calculation steps of the adaptive lattice interaction mechanism are as follows: Where, and are the high-frequency information and low-frequency information after interaction, α inter is the adaptive interaction coefficient, which is calculated by the strength of the local gradient inter , controls the interaction strength between high-frequency and low-frequency components, and the adaptive interaction coefficient α inter The specific calculation formula is: Where, β inter In order to control the sensitivity of the weighting coefficient to the local gradient, G local is the local gradient value, γ inter is the bias term, which controls the threshold of the gradient value; S32. Introducing a turbulence model into the multi-dimensional interaction enhancement module to simulate the interaction of features at different scales of the image, wherein the turbulence model is: Where, I mul is the multi-scale interaction feature, L is the number of scale layers, α l is the weighting coefficient for each scale, S l is the characteristic intensity related to the scale of the first layer, I l is the feature at the lth level, β l is the parameter that controls the intensity of the turbulence model, is the Laplace operator; S33. After completing the interaction enhancement of high frequency and low frequency and the interaction of different scale features, the different frequency and multi-scale interaction features are fused to obtain the multi-dimensional interaction feature I inter-mul .

5. The tower morphology monitoring method based on remote sensing image feature analysis according to claim 4 is characterized in that: In step S4, the specific steps of the adaptive segmentation framework are: S41. By introducing the local contrast enhancement mechanism in the cat's eye system, the important areas in the image are dynamically enhanced. The local areas in the tower remote sensing image are highlighted by an enhancement model based on local contrast. The contrast enhancement formula is: Where, I enha (x,y) is the pixel value after enhancement at position (x,y), I inter-mul (x,y) is the multidimensional interaction feature I inter-mul The pixel value at (x, y), I lmean (x, y) is the mean of the local area, I lstd (x, y) is the standard deviation of the local area; S42. Introduce the self-organized criticality theory to dynamically adjust the segmentation threshold. Automatically adjust the segmentation threshold according to the gradient change of the image. The segmentation threshold T(ti) is calculated as follows: Where, T init is the initial segmentation threshold, α T To control the rate of dynamic adjustment of the threshold, I lstd (x, y, ti) is the standard deviation of the local area at time ti, indicating the local variation of the image; S43, extracting the region of interest in the tower remote sensing image by calculating the relationship between the difference between the enhanced pixel value and the segmentation domain pixel value and the segmentation threshold, wherein the region boundary detection is: Where, if Region is 1, the pixel is the target area; if Region is 0, the pixel is the non-target area. seed The starting value of the segmentation domain.

6. The tower morphology monitoring method based on remote sensing image feature analysis according to claim 5 is characterized in that: In step S5, the tower shape monitoring model has the following specific steps: S51. First, the tower remote sensing image is input into the tower morphology monitoring model. The tower remote sensing image is preprocessed by removing environmental interference and sensor noise through adaptive noise suppression technology and multispectral fusion processing. Next, the adaptive weighted feature extraction method is used to dynamically weight each channel of the tower remote sensing image to extract the initial features of the tower remote sensing image, providing high-quality input features for subsequent tower remote sensing image feature extraction. S52. Inputting the initial features of the tower remote sensing image into M information interaction modules composed of a multi-frequency interference mechanism and a multi-dimensional interaction enhancement module, wherein the multi-frequency interference mechanism is used to perform dynamic frequency separation, and the separation of high-frequency information and low-frequency information is dynamically adjusted through phase superposition and interference of frequency components to obtain high-frequency information and low-frequency information of the tower remote sensing image. Subsequently, the multi-dimensional interaction enhancement module enhances the interaction between high-frequency information and low-frequency information and the fusion of multi-scale features by combining a lattice structure model and a turbulence model. Finally, the M interaction features are fused to obtain enhanced tower remote sensing image features. S53. After completing feature enhancement, the model dynamically adjusts the segmentation threshold through an adaptive segmentation framework, combined with the local contrast enhancement mechanism and self-organized criticality theory in the cat's eye system, to accurately extract the region of interest in the tower image. Finally, the extracted tower remote sensing area is input into the classifier for tower morphology classification. The tower status is classified according to different features in the area, including surface cracks, tilt and deformation, material aging, local deformation, and normal monitoring.

Images